2 * Performance counter core code
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/file.h>
17 #include <linux/poll.h>
18 #include <linux/sysfs.h>
19 #include <linux/ptrace.h>
20 #include <linux/percpu.h>
21 #include <linux/vmstat.h>
22 #include <linux/hardirq.h>
23 #include <linux/rculist.h>
24 #include <linux/uaccess.h>
25 #include <linux/syscalls.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/kernel_stat.h>
28 #include <linux/perf_counter.h>
29 #include <linux/dcache.h>
31 #include <asm/irq_regs.h>
34 * Each CPU has a list of per CPU counters:
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_tracking __read_mostly;
44 static atomic_t nr_munmap_tracking __read_mostly;
45 static atomic_t nr_comm_tracking __read_mostly;
47 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
48 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
51 * Lock for (sysadmin-configurable) counter reservations:
53 static DEFINE_SPINLOCK(perf_resource_lock);
56 * Architecture provided APIs - weak aliases:
58 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
63 void __weak hw_perf_disable(void) { barrier(); }
64 void __weak hw_perf_enable(void) { barrier(); }
66 void __weak hw_perf_counter_setup(int cpu) { barrier(); }
67 int __weak hw_perf_group_sched_in(struct perf_counter *group_leader,
68 struct perf_cpu_context *cpuctx,
69 struct perf_counter_context *ctx, int cpu)
74 void __weak perf_counter_print_debug(void) { }
76 static DEFINE_PER_CPU(int, disable_count);
78 void __perf_disable(void)
80 __get_cpu_var(disable_count)++;
83 bool __perf_enable(void)
85 return !--__get_cpu_var(disable_count);
88 void perf_disable(void)
94 void perf_enable(void)
101 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
103 struct perf_counter *group_leader = counter->group_leader;
106 * Depending on whether it is a standalone or sibling counter,
107 * add it straight to the context's counter list, or to the group
108 * leader's sibling list:
110 if (group_leader == counter)
111 list_add_tail(&counter->list_entry, &ctx->counter_list);
113 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
114 group_leader->nr_siblings++;
117 list_add_rcu(&counter->event_entry, &ctx->event_list);
122 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
124 struct perf_counter *sibling, *tmp;
128 list_del_init(&counter->list_entry);
129 list_del_rcu(&counter->event_entry);
131 if (counter->group_leader != counter)
132 counter->group_leader->nr_siblings--;
135 * If this was a group counter with sibling counters then
136 * upgrade the siblings to singleton counters by adding them
137 * to the context list directly:
139 list_for_each_entry_safe(sibling, tmp,
140 &counter->sibling_list, list_entry) {
142 list_move_tail(&sibling->list_entry, &ctx->counter_list);
143 sibling->group_leader = sibling;
148 counter_sched_out(struct perf_counter *counter,
149 struct perf_cpu_context *cpuctx,
150 struct perf_counter_context *ctx)
152 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
155 counter->state = PERF_COUNTER_STATE_INACTIVE;
156 counter->tstamp_stopped = ctx->time;
157 counter->pmu->disable(counter);
160 if (!is_software_counter(counter))
161 cpuctx->active_oncpu--;
163 if (counter->hw_event.exclusive || !cpuctx->active_oncpu)
164 cpuctx->exclusive = 0;
168 group_sched_out(struct perf_counter *group_counter,
169 struct perf_cpu_context *cpuctx,
170 struct perf_counter_context *ctx)
172 struct perf_counter *counter;
174 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
177 counter_sched_out(group_counter, cpuctx, ctx);
180 * Schedule out siblings (if any):
182 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
183 counter_sched_out(counter, cpuctx, ctx);
185 if (group_counter->hw_event.exclusive)
186 cpuctx->exclusive = 0;
190 * Cross CPU call to remove a performance counter
192 * We disable the counter on the hardware level first. After that we
193 * remove it from the context list.
195 static void __perf_counter_remove_from_context(void *info)
197 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
198 struct perf_counter *counter = info;
199 struct perf_counter_context *ctx = counter->ctx;
203 * If this is a task context, we need to check whether it is
204 * the current task context of this cpu. If not it has been
205 * scheduled out before the smp call arrived.
207 if (ctx->task && cpuctx->task_ctx != ctx)
210 spin_lock_irqsave(&ctx->lock, flags);
212 counter_sched_out(counter, cpuctx, ctx);
214 counter->task = NULL;
217 * Protect the list operation against NMI by disabling the
218 * counters on a global level. NOP for non NMI based counters.
221 list_del_counter(counter, ctx);
226 * Allow more per task counters with respect to the
229 cpuctx->max_pertask =
230 min(perf_max_counters - ctx->nr_counters,
231 perf_max_counters - perf_reserved_percpu);
234 spin_unlock_irqrestore(&ctx->lock, flags);
239 * Remove the counter from a task's (or a CPU's) list of counters.
241 * Must be called with counter->mutex and ctx->mutex held.
243 * CPU counters are removed with a smp call. For task counters we only
244 * call when the task is on a CPU.
246 static void perf_counter_remove_from_context(struct perf_counter *counter)
248 struct perf_counter_context *ctx = counter->ctx;
249 struct task_struct *task = ctx->task;
253 * Per cpu counters are removed via an smp call and
254 * the removal is always sucessful.
256 smp_call_function_single(counter->cpu,
257 __perf_counter_remove_from_context,
263 task_oncpu_function_call(task, __perf_counter_remove_from_context,
266 spin_lock_irq(&ctx->lock);
268 * If the context is active we need to retry the smp call.
270 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
271 spin_unlock_irq(&ctx->lock);
276 * The lock prevents that this context is scheduled in so we
277 * can remove the counter safely, if the call above did not
280 if (!list_empty(&counter->list_entry)) {
281 list_del_counter(counter, ctx);
282 counter->task = NULL;
284 spin_unlock_irq(&ctx->lock);
287 static inline u64 perf_clock(void)
289 return cpu_clock(smp_processor_id());
293 * Update the record of the current time in a context.
295 static void update_context_time(struct perf_counter_context *ctx)
297 u64 now = perf_clock();
299 ctx->time += now - ctx->timestamp;
300 ctx->timestamp = now;
304 * Update the total_time_enabled and total_time_running fields for a counter.
306 static void update_counter_times(struct perf_counter *counter)
308 struct perf_counter_context *ctx = counter->ctx;
311 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
314 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
316 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
317 run_end = counter->tstamp_stopped;
321 counter->total_time_running = run_end - counter->tstamp_running;
325 * Update total_time_enabled and total_time_running for all counters in a group.
327 static void update_group_times(struct perf_counter *leader)
329 struct perf_counter *counter;
331 update_counter_times(leader);
332 list_for_each_entry(counter, &leader->sibling_list, list_entry)
333 update_counter_times(counter);
337 * Cross CPU call to disable a performance counter
339 static void __perf_counter_disable(void *info)
341 struct perf_counter *counter = info;
342 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
343 struct perf_counter_context *ctx = counter->ctx;
347 * If this is a per-task counter, need to check whether this
348 * counter's task is the current task on this cpu.
350 if (ctx->task && cpuctx->task_ctx != ctx)
353 spin_lock_irqsave(&ctx->lock, flags);
356 * If the counter is on, turn it off.
357 * If it is in error state, leave it in error state.
359 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
360 update_context_time(ctx);
361 update_counter_times(counter);
362 if (counter == counter->group_leader)
363 group_sched_out(counter, cpuctx, ctx);
365 counter_sched_out(counter, cpuctx, ctx);
366 counter->state = PERF_COUNTER_STATE_OFF;
369 spin_unlock_irqrestore(&ctx->lock, flags);
375 static void perf_counter_disable(struct perf_counter *counter)
377 struct perf_counter_context *ctx = counter->ctx;
378 struct task_struct *task = ctx->task;
382 * Disable the counter on the cpu that it's on
384 smp_call_function_single(counter->cpu, __perf_counter_disable,
390 task_oncpu_function_call(task, __perf_counter_disable, counter);
392 spin_lock_irq(&ctx->lock);
394 * If the counter is still active, we need to retry the cross-call.
396 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
397 spin_unlock_irq(&ctx->lock);
402 * Since we have the lock this context can't be scheduled
403 * in, so we can change the state safely.
405 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
406 update_counter_times(counter);
407 counter->state = PERF_COUNTER_STATE_OFF;
410 spin_unlock_irq(&ctx->lock);
414 counter_sched_in(struct perf_counter *counter,
415 struct perf_cpu_context *cpuctx,
416 struct perf_counter_context *ctx,
419 if (counter->state <= PERF_COUNTER_STATE_OFF)
422 counter->state = PERF_COUNTER_STATE_ACTIVE;
423 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
425 * The new state must be visible before we turn it on in the hardware:
429 if (counter->pmu->enable(counter)) {
430 counter->state = PERF_COUNTER_STATE_INACTIVE;
435 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
437 if (!is_software_counter(counter))
438 cpuctx->active_oncpu++;
441 if (counter->hw_event.exclusive)
442 cpuctx->exclusive = 1;
448 group_sched_in(struct perf_counter *group_counter,
449 struct perf_cpu_context *cpuctx,
450 struct perf_counter_context *ctx,
453 struct perf_counter *counter, *partial_group;
456 if (group_counter->state == PERF_COUNTER_STATE_OFF)
459 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
461 return ret < 0 ? ret : 0;
463 group_counter->prev_state = group_counter->state;
464 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
468 * Schedule in siblings as one group (if any):
470 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
471 counter->prev_state = counter->state;
472 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
473 partial_group = counter;
482 * Groups can be scheduled in as one unit only, so undo any
483 * partial group before returning:
485 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
486 if (counter == partial_group)
488 counter_sched_out(counter, cpuctx, ctx);
490 counter_sched_out(group_counter, cpuctx, ctx);
496 * Return 1 for a group consisting entirely of software counters,
497 * 0 if the group contains any hardware counters.
499 static int is_software_only_group(struct perf_counter *leader)
501 struct perf_counter *counter;
503 if (!is_software_counter(leader))
506 list_for_each_entry(counter, &leader->sibling_list, list_entry)
507 if (!is_software_counter(counter))
514 * Work out whether we can put this counter group on the CPU now.
516 static int group_can_go_on(struct perf_counter *counter,
517 struct perf_cpu_context *cpuctx,
521 * Groups consisting entirely of software counters can always go on.
523 if (is_software_only_group(counter))
526 * If an exclusive group is already on, no other hardware
527 * counters can go on.
529 if (cpuctx->exclusive)
532 * If this group is exclusive and there are already
533 * counters on the CPU, it can't go on.
535 if (counter->hw_event.exclusive && cpuctx->active_oncpu)
538 * Otherwise, try to add it if all previous groups were able
544 static void add_counter_to_ctx(struct perf_counter *counter,
545 struct perf_counter_context *ctx)
547 list_add_counter(counter, ctx);
548 counter->prev_state = PERF_COUNTER_STATE_OFF;
549 counter->tstamp_enabled = ctx->time;
550 counter->tstamp_running = ctx->time;
551 counter->tstamp_stopped = ctx->time;
555 * Cross CPU call to install and enable a performance counter
557 static void __perf_install_in_context(void *info)
559 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
560 struct perf_counter *counter = info;
561 struct perf_counter_context *ctx = counter->ctx;
562 struct perf_counter *leader = counter->group_leader;
563 int cpu = smp_processor_id();
568 * If this is a task context, we need to check whether it is
569 * the current task context of this cpu. If not it has been
570 * scheduled out before the smp call arrived.
572 if (ctx->task && cpuctx->task_ctx != ctx)
575 spin_lock_irqsave(&ctx->lock, flags);
576 update_context_time(ctx);
579 * Protect the list operation against NMI by disabling the
580 * counters on a global level. NOP for non NMI based counters.
584 add_counter_to_ctx(counter, ctx);
587 * Don't put the counter on if it is disabled or if
588 * it is in a group and the group isn't on.
590 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
591 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
595 * An exclusive counter can't go on if there are already active
596 * hardware counters, and no hardware counter can go on if there
597 * is already an exclusive counter on.
599 if (!group_can_go_on(counter, cpuctx, 1))
602 err = counter_sched_in(counter, cpuctx, ctx, cpu);
606 * This counter couldn't go on. If it is in a group
607 * then we have to pull the whole group off.
608 * If the counter group is pinned then put it in error state.
610 if (leader != counter)
611 group_sched_out(leader, cpuctx, ctx);
612 if (leader->hw_event.pinned) {
613 update_group_times(leader);
614 leader->state = PERF_COUNTER_STATE_ERROR;
618 if (!err && !ctx->task && cpuctx->max_pertask)
619 cpuctx->max_pertask--;
624 spin_unlock_irqrestore(&ctx->lock, flags);
628 * Attach a performance counter to a context
630 * First we add the counter to the list with the hardware enable bit
631 * in counter->hw_config cleared.
633 * If the counter is attached to a task which is on a CPU we use a smp
634 * call to enable it in the task context. The task might have been
635 * scheduled away, but we check this in the smp call again.
637 * Must be called with ctx->mutex held.
640 perf_install_in_context(struct perf_counter_context *ctx,
641 struct perf_counter *counter,
644 struct task_struct *task = ctx->task;
648 * Per cpu counters are installed via an smp call and
649 * the install is always sucessful.
651 smp_call_function_single(cpu, __perf_install_in_context,
656 counter->task = task;
658 task_oncpu_function_call(task, __perf_install_in_context,
661 spin_lock_irq(&ctx->lock);
663 * we need to retry the smp call.
665 if (ctx->is_active && list_empty(&counter->list_entry)) {
666 spin_unlock_irq(&ctx->lock);
671 * The lock prevents that this context is scheduled in so we
672 * can add the counter safely, if it the call above did not
675 if (list_empty(&counter->list_entry))
676 add_counter_to_ctx(counter, ctx);
677 spin_unlock_irq(&ctx->lock);
681 * Cross CPU call to enable a performance counter
683 static void __perf_counter_enable(void *info)
685 struct perf_counter *counter = info;
686 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
687 struct perf_counter_context *ctx = counter->ctx;
688 struct perf_counter *leader = counter->group_leader;
693 * If this is a per-task counter, need to check whether this
694 * counter's task is the current task on this cpu.
696 if (ctx->task && cpuctx->task_ctx != ctx)
699 spin_lock_irqsave(&ctx->lock, flags);
700 update_context_time(ctx);
702 counter->prev_state = counter->state;
703 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
705 counter->state = PERF_COUNTER_STATE_INACTIVE;
706 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
709 * If the counter is in a group and isn't the group leader,
710 * then don't put it on unless the group is on.
712 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
715 if (!group_can_go_on(counter, cpuctx, 1)) {
719 if (counter == leader)
720 err = group_sched_in(counter, cpuctx, ctx,
723 err = counter_sched_in(counter, cpuctx, ctx,
730 * If this counter can't go on and it's part of a
731 * group, then the whole group has to come off.
733 if (leader != counter)
734 group_sched_out(leader, cpuctx, ctx);
735 if (leader->hw_event.pinned) {
736 update_group_times(leader);
737 leader->state = PERF_COUNTER_STATE_ERROR;
742 spin_unlock_irqrestore(&ctx->lock, flags);
748 static void perf_counter_enable(struct perf_counter *counter)
750 struct perf_counter_context *ctx = counter->ctx;
751 struct task_struct *task = ctx->task;
755 * Enable the counter on the cpu that it's on
757 smp_call_function_single(counter->cpu, __perf_counter_enable,
762 spin_lock_irq(&ctx->lock);
763 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
767 * If the counter is in error state, clear that first.
768 * That way, if we see the counter in error state below, we
769 * know that it has gone back into error state, as distinct
770 * from the task having been scheduled away before the
771 * cross-call arrived.
773 if (counter->state == PERF_COUNTER_STATE_ERROR)
774 counter->state = PERF_COUNTER_STATE_OFF;
777 spin_unlock_irq(&ctx->lock);
778 task_oncpu_function_call(task, __perf_counter_enable, counter);
780 spin_lock_irq(&ctx->lock);
783 * If the context is active and the counter is still off,
784 * we need to retry the cross-call.
786 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
790 * Since we have the lock this context can't be scheduled
791 * in, so we can change the state safely.
793 if (counter->state == PERF_COUNTER_STATE_OFF) {
794 counter->state = PERF_COUNTER_STATE_INACTIVE;
795 counter->tstamp_enabled =
796 ctx->time - counter->total_time_enabled;
799 spin_unlock_irq(&ctx->lock);
802 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
805 * not supported on inherited counters
807 if (counter->hw_event.inherit)
810 atomic_add(refresh, &counter->event_limit);
811 perf_counter_enable(counter);
816 void __perf_counter_sched_out(struct perf_counter_context *ctx,
817 struct perf_cpu_context *cpuctx)
819 struct perf_counter *counter;
821 spin_lock(&ctx->lock);
823 if (likely(!ctx->nr_counters))
825 update_context_time(ctx);
828 if (ctx->nr_active) {
829 list_for_each_entry(counter, &ctx->counter_list, list_entry)
830 group_sched_out(counter, cpuctx, ctx);
834 spin_unlock(&ctx->lock);
838 * Called from scheduler to remove the counters of the current task,
839 * with interrupts disabled.
841 * We stop each counter and update the counter value in counter->count.
843 * This does not protect us against NMI, but disable()
844 * sets the disabled bit in the control field of counter _before_
845 * accessing the counter control register. If a NMI hits, then it will
846 * not restart the counter.
848 void perf_counter_task_sched_out(struct task_struct *task, int cpu)
850 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
851 struct perf_counter_context *ctx = &task->perf_counter_ctx;
852 struct pt_regs *regs;
854 if (likely(!cpuctx->task_ctx))
857 update_context_time(ctx);
859 regs = task_pt_regs(task);
860 perf_swcounter_event(PERF_COUNT_CONTEXT_SWITCHES, 1, 1, regs, 0);
861 __perf_counter_sched_out(ctx, cpuctx);
863 cpuctx->task_ctx = NULL;
866 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
868 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
870 __perf_counter_sched_out(ctx, cpuctx);
871 cpuctx->task_ctx = NULL;
874 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
876 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
880 __perf_counter_sched_in(struct perf_counter_context *ctx,
881 struct perf_cpu_context *cpuctx, int cpu)
883 struct perf_counter *counter;
886 spin_lock(&ctx->lock);
888 if (likely(!ctx->nr_counters))
891 ctx->timestamp = perf_clock();
896 * First go through the list and put on any pinned groups
897 * in order to give them the best chance of going on.
899 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
900 if (counter->state <= PERF_COUNTER_STATE_OFF ||
901 !counter->hw_event.pinned)
903 if (counter->cpu != -1 && counter->cpu != cpu)
906 if (group_can_go_on(counter, cpuctx, 1))
907 group_sched_in(counter, cpuctx, ctx, cpu);
910 * If this pinned group hasn't been scheduled,
911 * put it in error state.
913 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
914 update_group_times(counter);
915 counter->state = PERF_COUNTER_STATE_ERROR;
919 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
921 * Ignore counters in OFF or ERROR state, and
922 * ignore pinned counters since we did them already.
924 if (counter->state <= PERF_COUNTER_STATE_OFF ||
925 counter->hw_event.pinned)
929 * Listen to the 'cpu' scheduling filter constraint
932 if (counter->cpu != -1 && counter->cpu != cpu)
935 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
936 if (group_sched_in(counter, cpuctx, ctx, cpu))
942 spin_unlock(&ctx->lock);
946 * Called from scheduler to add the counters of the current task
947 * with interrupts disabled.
949 * We restore the counter value and then enable it.
951 * This does not protect us against NMI, but enable()
952 * sets the enabled bit in the control field of counter _before_
953 * accessing the counter control register. If a NMI hits, then it will
954 * keep the counter running.
956 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
958 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
959 struct perf_counter_context *ctx = &task->perf_counter_ctx;
961 __perf_counter_sched_in(ctx, cpuctx, cpu);
962 cpuctx->task_ctx = ctx;
965 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
967 struct perf_counter_context *ctx = &cpuctx->ctx;
969 __perf_counter_sched_in(ctx, cpuctx, cpu);
972 int perf_counter_task_disable(void)
974 struct task_struct *curr = current;
975 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
976 struct perf_counter *counter;
979 if (likely(!ctx->nr_counters))
982 local_irq_save(flags);
984 __perf_counter_task_sched_out(ctx);
986 spin_lock(&ctx->lock);
989 * Disable all the counters:
993 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
994 if (counter->state != PERF_COUNTER_STATE_ERROR) {
995 update_group_times(counter);
996 counter->state = PERF_COUNTER_STATE_OFF;
1002 spin_unlock_irqrestore(&ctx->lock, flags);
1007 int perf_counter_task_enable(void)
1009 struct task_struct *curr = current;
1010 struct perf_counter_context *ctx = &curr->perf_counter_ctx;
1011 struct perf_counter *counter;
1012 unsigned long flags;
1015 if (likely(!ctx->nr_counters))
1018 local_irq_save(flags);
1019 cpu = smp_processor_id();
1021 __perf_counter_task_sched_out(ctx);
1023 spin_lock(&ctx->lock);
1026 * Disable all the counters:
1030 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1031 if (counter->state > PERF_COUNTER_STATE_OFF)
1033 counter->state = PERF_COUNTER_STATE_INACTIVE;
1034 counter->tstamp_enabled =
1035 ctx->time - counter->total_time_enabled;
1036 counter->hw_event.disabled = 0;
1040 spin_unlock(&ctx->lock);
1042 perf_counter_task_sched_in(curr, cpu);
1044 local_irq_restore(flags);
1049 static void perf_log_period(struct perf_counter *counter, u64 period);
1051 static void perf_adjust_freq(struct perf_counter_context *ctx)
1053 struct perf_counter *counter;
1058 spin_lock(&ctx->lock);
1059 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1060 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1063 if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
1066 events = HZ * counter->hw.interrupts * counter->hw.irq_period;
1067 period = div64_u64(events, counter->hw_event.irq_freq);
1069 delta = (s64)(1 + period - counter->hw.irq_period);
1072 irq_period = counter->hw.irq_period + delta;
1077 perf_log_period(counter, irq_period);
1079 counter->hw.irq_period = irq_period;
1080 counter->hw.interrupts = 0;
1082 spin_unlock(&ctx->lock);
1086 * Round-robin a context's counters:
1088 static void rotate_ctx(struct perf_counter_context *ctx)
1090 struct perf_counter *counter;
1092 if (!ctx->nr_counters)
1095 spin_lock(&ctx->lock);
1097 * Rotate the first entry last (works just fine for group counters too):
1100 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1101 list_move_tail(&counter->list_entry, &ctx->counter_list);
1106 spin_unlock(&ctx->lock);
1109 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1111 struct perf_cpu_context *cpuctx;
1112 struct perf_counter_context *ctx;
1114 if (!atomic_read(&nr_counters))
1117 cpuctx = &per_cpu(perf_cpu_context, cpu);
1118 ctx = &curr->perf_counter_ctx;
1120 perf_adjust_freq(&cpuctx->ctx);
1121 perf_adjust_freq(ctx);
1123 perf_counter_cpu_sched_out(cpuctx);
1124 __perf_counter_task_sched_out(ctx);
1126 rotate_ctx(&cpuctx->ctx);
1129 perf_counter_cpu_sched_in(cpuctx, cpu);
1130 perf_counter_task_sched_in(curr, cpu);
1134 * Cross CPU call to read the hardware counter
1136 static void __read(void *info)
1138 struct perf_counter *counter = info;
1139 struct perf_counter_context *ctx = counter->ctx;
1140 unsigned long flags;
1142 local_irq_save(flags);
1144 update_context_time(ctx);
1145 counter->pmu->read(counter);
1146 update_counter_times(counter);
1147 local_irq_restore(flags);
1150 static u64 perf_counter_read(struct perf_counter *counter)
1153 * If counter is enabled and currently active on a CPU, update the
1154 * value in the counter structure:
1156 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1157 smp_call_function_single(counter->oncpu,
1158 __read, counter, 1);
1159 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1160 update_counter_times(counter);
1163 return atomic64_read(&counter->count);
1166 static void put_context(struct perf_counter_context *ctx)
1169 put_task_struct(ctx->task);
1172 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1174 struct perf_cpu_context *cpuctx;
1175 struct perf_counter_context *ctx;
1176 struct task_struct *task;
1179 * If cpu is not a wildcard then this is a percpu counter:
1182 /* Must be root to operate on a CPU counter: */
1183 if (sysctl_perf_counter_priv && !capable(CAP_SYS_ADMIN))
1184 return ERR_PTR(-EACCES);
1186 if (cpu < 0 || cpu > num_possible_cpus())
1187 return ERR_PTR(-EINVAL);
1190 * We could be clever and allow to attach a counter to an
1191 * offline CPU and activate it when the CPU comes up, but
1194 if (!cpu_isset(cpu, cpu_online_map))
1195 return ERR_PTR(-ENODEV);
1197 cpuctx = &per_cpu(perf_cpu_context, cpu);
1207 task = find_task_by_vpid(pid);
1209 get_task_struct(task);
1213 return ERR_PTR(-ESRCH);
1215 ctx = &task->perf_counter_ctx;
1218 /* Reuse ptrace permission checks for now. */
1219 if (!ptrace_may_access(task, PTRACE_MODE_READ)) {
1221 return ERR_PTR(-EACCES);
1227 static void free_counter_rcu(struct rcu_head *head)
1229 struct perf_counter *counter;
1231 counter = container_of(head, struct perf_counter, rcu_head);
1235 static void perf_pending_sync(struct perf_counter *counter);
1237 static void free_counter(struct perf_counter *counter)
1239 perf_pending_sync(counter);
1241 atomic_dec(&nr_counters);
1242 if (counter->hw_event.mmap)
1243 atomic_dec(&nr_mmap_tracking);
1244 if (counter->hw_event.munmap)
1245 atomic_dec(&nr_munmap_tracking);
1246 if (counter->hw_event.comm)
1247 atomic_dec(&nr_comm_tracking);
1249 if (counter->destroy)
1250 counter->destroy(counter);
1252 call_rcu(&counter->rcu_head, free_counter_rcu);
1256 * Called when the last reference to the file is gone.
1258 static int perf_release(struct inode *inode, struct file *file)
1260 struct perf_counter *counter = file->private_data;
1261 struct perf_counter_context *ctx = counter->ctx;
1263 file->private_data = NULL;
1265 mutex_lock(&ctx->mutex);
1266 mutex_lock(&counter->mutex);
1268 perf_counter_remove_from_context(counter);
1270 mutex_unlock(&counter->mutex);
1271 mutex_unlock(&ctx->mutex);
1273 free_counter(counter);
1280 * Read the performance counter - simple non blocking version for now
1283 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1289 * Return end-of-file for a read on a counter that is in
1290 * error state (i.e. because it was pinned but it couldn't be
1291 * scheduled on to the CPU at some point).
1293 if (counter->state == PERF_COUNTER_STATE_ERROR)
1296 mutex_lock(&counter->mutex);
1297 values[0] = perf_counter_read(counter);
1299 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1300 values[n++] = counter->total_time_enabled +
1301 atomic64_read(&counter->child_total_time_enabled);
1302 if (counter->hw_event.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1303 values[n++] = counter->total_time_running +
1304 atomic64_read(&counter->child_total_time_running);
1305 mutex_unlock(&counter->mutex);
1307 if (count < n * sizeof(u64))
1309 count = n * sizeof(u64);
1311 if (copy_to_user(buf, values, count))
1318 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1320 struct perf_counter *counter = file->private_data;
1322 return perf_read_hw(counter, buf, count);
1325 static unsigned int perf_poll(struct file *file, poll_table *wait)
1327 struct perf_counter *counter = file->private_data;
1328 struct perf_mmap_data *data;
1329 unsigned int events = POLL_HUP;
1332 data = rcu_dereference(counter->data);
1334 events = atomic_xchg(&data->poll, 0);
1337 poll_wait(file, &counter->waitq, wait);
1342 static void perf_counter_reset(struct perf_counter *counter)
1344 (void)perf_counter_read(counter);
1345 atomic64_set(&counter->count, 0);
1346 perf_counter_update_userpage(counter);
1349 static void perf_counter_for_each_sibling(struct perf_counter *counter,
1350 void (*func)(struct perf_counter *))
1352 struct perf_counter_context *ctx = counter->ctx;
1353 struct perf_counter *sibling;
1355 spin_lock_irq(&ctx->lock);
1356 counter = counter->group_leader;
1359 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1361 spin_unlock_irq(&ctx->lock);
1364 static void perf_counter_for_each_child(struct perf_counter *counter,
1365 void (*func)(struct perf_counter *))
1367 struct perf_counter *child;
1369 mutex_lock(&counter->mutex);
1371 list_for_each_entry(child, &counter->child_list, child_list)
1373 mutex_unlock(&counter->mutex);
1376 static void perf_counter_for_each(struct perf_counter *counter,
1377 void (*func)(struct perf_counter *))
1379 struct perf_counter *child;
1381 mutex_lock(&counter->mutex);
1382 perf_counter_for_each_sibling(counter, func);
1383 list_for_each_entry(child, &counter->child_list, child_list)
1384 perf_counter_for_each_sibling(child, func);
1385 mutex_unlock(&counter->mutex);
1388 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1390 struct perf_counter *counter = file->private_data;
1391 void (*func)(struct perf_counter *);
1395 case PERF_COUNTER_IOC_ENABLE:
1396 func = perf_counter_enable;
1398 case PERF_COUNTER_IOC_DISABLE:
1399 func = perf_counter_disable;
1401 case PERF_COUNTER_IOC_RESET:
1402 func = perf_counter_reset;
1405 case PERF_COUNTER_IOC_REFRESH:
1406 return perf_counter_refresh(counter, arg);
1411 if (flags & PERF_IOC_FLAG_GROUP)
1412 perf_counter_for_each(counter, func);
1414 perf_counter_for_each_child(counter, func);
1420 * Callers need to ensure there can be no nesting of this function, otherwise
1421 * the seqlock logic goes bad. We can not serialize this because the arch
1422 * code calls this from NMI context.
1424 void perf_counter_update_userpage(struct perf_counter *counter)
1426 struct perf_mmap_data *data;
1427 struct perf_counter_mmap_page *userpg;
1430 data = rcu_dereference(counter->data);
1434 userpg = data->user_page;
1437 * Disable preemption so as to not let the corresponding user-space
1438 * spin too long if we get preempted.
1443 userpg->index = counter->hw.idx;
1444 userpg->offset = atomic64_read(&counter->count);
1445 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1446 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1455 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1457 struct perf_counter *counter = vma->vm_file->private_data;
1458 struct perf_mmap_data *data;
1459 int ret = VM_FAULT_SIGBUS;
1462 data = rcu_dereference(counter->data);
1466 if (vmf->pgoff == 0) {
1467 vmf->page = virt_to_page(data->user_page);
1469 int nr = vmf->pgoff - 1;
1471 if ((unsigned)nr > data->nr_pages)
1474 vmf->page = virt_to_page(data->data_pages[nr]);
1476 get_page(vmf->page);
1484 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1486 struct perf_mmap_data *data;
1490 WARN_ON(atomic_read(&counter->mmap_count));
1492 size = sizeof(struct perf_mmap_data);
1493 size += nr_pages * sizeof(void *);
1495 data = kzalloc(size, GFP_KERNEL);
1499 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1500 if (!data->user_page)
1501 goto fail_user_page;
1503 for (i = 0; i < nr_pages; i++) {
1504 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1505 if (!data->data_pages[i])
1506 goto fail_data_pages;
1509 data->nr_pages = nr_pages;
1510 atomic_set(&data->lock, -1);
1512 rcu_assign_pointer(counter->data, data);
1517 for (i--; i >= 0; i--)
1518 free_page((unsigned long)data->data_pages[i]);
1520 free_page((unsigned long)data->user_page);
1529 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
1531 struct perf_mmap_data *data = container_of(rcu_head,
1532 struct perf_mmap_data, rcu_head);
1535 free_page((unsigned long)data->user_page);
1536 for (i = 0; i < data->nr_pages; i++)
1537 free_page((unsigned long)data->data_pages[i]);
1541 static void perf_mmap_data_free(struct perf_counter *counter)
1543 struct perf_mmap_data *data = counter->data;
1545 WARN_ON(atomic_read(&counter->mmap_count));
1547 rcu_assign_pointer(counter->data, NULL);
1548 call_rcu(&data->rcu_head, __perf_mmap_data_free);
1551 static void perf_mmap_open(struct vm_area_struct *vma)
1553 struct perf_counter *counter = vma->vm_file->private_data;
1555 atomic_inc(&counter->mmap_count);
1558 static void perf_mmap_close(struct vm_area_struct *vma)
1560 struct perf_counter *counter = vma->vm_file->private_data;
1562 if (atomic_dec_and_mutex_lock(&counter->mmap_count,
1563 &counter->mmap_mutex)) {
1564 struct user_struct *user = current_user();
1566 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
1567 vma->vm_mm->locked_vm -= counter->data->nr_locked;
1568 perf_mmap_data_free(counter);
1569 mutex_unlock(&counter->mmap_mutex);
1573 static struct vm_operations_struct perf_mmap_vmops = {
1574 .open = perf_mmap_open,
1575 .close = perf_mmap_close,
1576 .fault = perf_mmap_fault,
1579 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
1581 struct perf_counter *counter = file->private_data;
1582 struct user_struct *user = current_user();
1583 unsigned long vma_size;
1584 unsigned long nr_pages;
1585 unsigned long user_locked, user_lock_limit;
1586 unsigned long locked, lock_limit;
1587 long user_extra, extra;
1590 if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
1593 vma_size = vma->vm_end - vma->vm_start;
1594 nr_pages = (vma_size / PAGE_SIZE) - 1;
1597 * If we have data pages ensure they're a power-of-two number, so we
1598 * can do bitmasks instead of modulo.
1600 if (nr_pages != 0 && !is_power_of_2(nr_pages))
1603 if (vma_size != PAGE_SIZE * (1 + nr_pages))
1606 if (vma->vm_pgoff != 0)
1609 mutex_lock(&counter->mmap_mutex);
1610 if (atomic_inc_not_zero(&counter->mmap_count)) {
1611 if (nr_pages != counter->data->nr_pages)
1616 user_extra = nr_pages + 1;
1617 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
1618 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
1621 if (user_locked > user_lock_limit)
1622 extra = user_locked - user_lock_limit;
1624 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1625 lock_limit >>= PAGE_SHIFT;
1626 locked = vma->vm_mm->locked_vm + extra;
1628 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
1633 WARN_ON(counter->data);
1634 ret = perf_mmap_data_alloc(counter, nr_pages);
1638 atomic_set(&counter->mmap_count, 1);
1639 atomic_long_add(user_extra, &user->locked_vm);
1640 vma->vm_mm->locked_vm += extra;
1641 counter->data->nr_locked = extra;
1643 mutex_unlock(&counter->mmap_mutex);
1645 vma->vm_flags &= ~VM_MAYWRITE;
1646 vma->vm_flags |= VM_RESERVED;
1647 vma->vm_ops = &perf_mmap_vmops;
1652 static int perf_fasync(int fd, struct file *filp, int on)
1654 struct perf_counter *counter = filp->private_data;
1655 struct inode *inode = filp->f_path.dentry->d_inode;
1658 mutex_lock(&inode->i_mutex);
1659 retval = fasync_helper(fd, filp, on, &counter->fasync);
1660 mutex_unlock(&inode->i_mutex);
1668 static const struct file_operations perf_fops = {
1669 .release = perf_release,
1672 .unlocked_ioctl = perf_ioctl,
1673 .compat_ioctl = perf_ioctl,
1675 .fasync = perf_fasync,
1679 * Perf counter wakeup
1681 * If there's data, ensure we set the poll() state and publish everything
1682 * to user-space before waking everybody up.
1685 void perf_counter_wakeup(struct perf_counter *counter)
1687 wake_up_all(&counter->waitq);
1689 if (counter->pending_kill) {
1690 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
1691 counter->pending_kill = 0;
1698 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
1700 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
1701 * single linked list and use cmpxchg() to add entries lockless.
1704 static void perf_pending_counter(struct perf_pending_entry *entry)
1706 struct perf_counter *counter = container_of(entry,
1707 struct perf_counter, pending);
1709 if (counter->pending_disable) {
1710 counter->pending_disable = 0;
1711 perf_counter_disable(counter);
1714 if (counter->pending_wakeup) {
1715 counter->pending_wakeup = 0;
1716 perf_counter_wakeup(counter);
1720 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
1722 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
1726 static void perf_pending_queue(struct perf_pending_entry *entry,
1727 void (*func)(struct perf_pending_entry *))
1729 struct perf_pending_entry **head;
1731 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
1736 head = &get_cpu_var(perf_pending_head);
1739 entry->next = *head;
1740 } while (cmpxchg(head, entry->next, entry) != entry->next);
1742 set_perf_counter_pending();
1744 put_cpu_var(perf_pending_head);
1747 static int __perf_pending_run(void)
1749 struct perf_pending_entry *list;
1752 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
1753 while (list != PENDING_TAIL) {
1754 void (*func)(struct perf_pending_entry *);
1755 struct perf_pending_entry *entry = list;
1762 * Ensure we observe the unqueue before we issue the wakeup,
1763 * so that we won't be waiting forever.
1764 * -- see perf_not_pending().
1775 static inline int perf_not_pending(struct perf_counter *counter)
1778 * If we flush on whatever cpu we run, there is a chance we don't
1782 __perf_pending_run();
1786 * Ensure we see the proper queue state before going to sleep
1787 * so that we do not miss the wakeup. -- see perf_pending_handle()
1790 return counter->pending.next == NULL;
1793 static void perf_pending_sync(struct perf_counter *counter)
1795 wait_event(counter->waitq, perf_not_pending(counter));
1798 void perf_counter_do_pending(void)
1800 __perf_pending_run();
1804 * Callchain support -- arch specific
1807 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1816 struct perf_output_handle {
1817 struct perf_counter *counter;
1818 struct perf_mmap_data *data;
1819 unsigned int offset;
1824 unsigned long flags;
1827 static void perf_output_wakeup(struct perf_output_handle *handle)
1829 atomic_set(&handle->data->poll, POLL_IN);
1832 handle->counter->pending_wakeup = 1;
1833 perf_pending_queue(&handle->counter->pending,
1834 perf_pending_counter);
1836 perf_counter_wakeup(handle->counter);
1840 * Curious locking construct.
1842 * We need to ensure a later event doesn't publish a head when a former
1843 * event isn't done writing. However since we need to deal with NMIs we
1844 * cannot fully serialize things.
1846 * What we do is serialize between CPUs so we only have to deal with NMI
1847 * nesting on a single CPU.
1849 * We only publish the head (and generate a wakeup) when the outer-most
1852 static void perf_output_lock(struct perf_output_handle *handle)
1854 struct perf_mmap_data *data = handle->data;
1859 local_irq_save(handle->flags);
1860 cpu = smp_processor_id();
1862 if (in_nmi() && atomic_read(&data->lock) == cpu)
1865 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1871 static void perf_output_unlock(struct perf_output_handle *handle)
1873 struct perf_mmap_data *data = handle->data;
1876 data->done_head = data->head;
1878 if (!handle->locked)
1883 * The xchg implies a full barrier that ensures all writes are done
1884 * before we publish the new head, matched by a rmb() in userspace when
1885 * reading this position.
1887 while ((head = atomic_xchg(&data->done_head, 0)))
1888 data->user_page->data_head = head;
1891 * NMI can happen here, which means we can miss a done_head update.
1894 cpu = atomic_xchg(&data->lock, -1);
1895 WARN_ON_ONCE(cpu != smp_processor_id());
1898 * Therefore we have to validate we did not indeed do so.
1900 if (unlikely(atomic_read(&data->done_head))) {
1902 * Since we had it locked, we can lock it again.
1904 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
1910 if (atomic_xchg(&data->wakeup, 0))
1911 perf_output_wakeup(handle);
1913 local_irq_restore(handle->flags);
1916 static int perf_output_begin(struct perf_output_handle *handle,
1917 struct perf_counter *counter, unsigned int size,
1918 int nmi, int overflow)
1920 struct perf_mmap_data *data;
1921 unsigned int offset, head;
1924 * For inherited counters we send all the output towards the parent.
1926 if (counter->parent)
1927 counter = counter->parent;
1930 data = rcu_dereference(counter->data);
1934 handle->data = data;
1935 handle->counter = counter;
1937 handle->overflow = overflow;
1939 if (!data->nr_pages)
1942 perf_output_lock(handle);
1945 offset = head = atomic_read(&data->head);
1947 } while (atomic_cmpxchg(&data->head, offset, head) != offset);
1949 handle->offset = offset;
1950 handle->head = head;
1952 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
1953 atomic_set(&data->wakeup, 1);
1958 perf_output_wakeup(handle);
1965 static void perf_output_copy(struct perf_output_handle *handle,
1966 void *buf, unsigned int len)
1968 unsigned int pages_mask;
1969 unsigned int offset;
1973 offset = handle->offset;
1974 pages_mask = handle->data->nr_pages - 1;
1975 pages = handle->data->data_pages;
1978 unsigned int page_offset;
1981 nr = (offset >> PAGE_SHIFT) & pages_mask;
1982 page_offset = offset & (PAGE_SIZE - 1);
1983 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
1985 memcpy(pages[nr] + page_offset, buf, size);
1992 handle->offset = offset;
1995 * Check we didn't copy past our reservation window, taking the
1996 * possible unsigned int wrap into account.
1998 WARN_ON_ONCE(((int)(handle->head - handle->offset)) < 0);
2001 #define perf_output_put(handle, x) \
2002 perf_output_copy((handle), &(x), sizeof(x))
2004 static void perf_output_end(struct perf_output_handle *handle)
2006 struct perf_counter *counter = handle->counter;
2007 struct perf_mmap_data *data = handle->data;
2009 int wakeup_events = counter->hw_event.wakeup_events;
2011 if (handle->overflow && wakeup_events) {
2012 int events = atomic_inc_return(&data->events);
2013 if (events >= wakeup_events) {
2014 atomic_sub(wakeup_events, &data->events);
2015 atomic_set(&data->wakeup, 1);
2019 perf_output_unlock(handle);
2023 static void perf_counter_output(struct perf_counter *counter,
2024 int nmi, struct pt_regs *regs, u64 addr)
2027 u64 record_type = counter->hw_event.record_type;
2028 struct perf_output_handle handle;
2029 struct perf_event_header header;
2038 struct perf_callchain_entry *callchain = NULL;
2039 int callchain_size = 0;
2046 header.size = sizeof(header);
2048 header.misc = PERF_EVENT_MISC_OVERFLOW;
2049 header.misc |= perf_misc_flags(regs);
2051 if (record_type & PERF_RECORD_IP) {
2052 ip = perf_instruction_pointer(regs);
2053 header.type |= PERF_RECORD_IP;
2054 header.size += sizeof(ip);
2057 if (record_type & PERF_RECORD_TID) {
2058 /* namespace issues */
2059 tid_entry.pid = current->group_leader->pid;
2060 tid_entry.tid = current->pid;
2062 header.type |= PERF_RECORD_TID;
2063 header.size += sizeof(tid_entry);
2066 if (record_type & PERF_RECORD_TIME) {
2068 * Maybe do better on x86 and provide cpu_clock_nmi()
2070 time = sched_clock();
2072 header.type |= PERF_RECORD_TIME;
2073 header.size += sizeof(u64);
2076 if (record_type & PERF_RECORD_ADDR) {
2077 header.type |= PERF_RECORD_ADDR;
2078 header.size += sizeof(u64);
2081 if (record_type & PERF_RECORD_CONFIG) {
2082 header.type |= PERF_RECORD_CONFIG;
2083 header.size += sizeof(u64);
2086 if (record_type & PERF_RECORD_CPU) {
2087 header.type |= PERF_RECORD_CPU;
2088 header.size += sizeof(cpu_entry);
2090 cpu_entry.cpu = raw_smp_processor_id();
2093 if (record_type & PERF_RECORD_GROUP) {
2094 header.type |= PERF_RECORD_GROUP;
2095 header.size += sizeof(u64) +
2096 counter->nr_siblings * sizeof(group_entry);
2099 if (record_type & PERF_RECORD_CALLCHAIN) {
2100 callchain = perf_callchain(regs);
2103 callchain_size = (1 + callchain->nr) * sizeof(u64);
2105 header.type |= PERF_RECORD_CALLCHAIN;
2106 header.size += callchain_size;
2110 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2114 perf_output_put(&handle, header);
2116 if (record_type & PERF_RECORD_IP)
2117 perf_output_put(&handle, ip);
2119 if (record_type & PERF_RECORD_TID)
2120 perf_output_put(&handle, tid_entry);
2122 if (record_type & PERF_RECORD_TIME)
2123 perf_output_put(&handle, time);
2125 if (record_type & PERF_RECORD_ADDR)
2126 perf_output_put(&handle, addr);
2128 if (record_type & PERF_RECORD_CONFIG)
2129 perf_output_put(&handle, counter->hw_event.config);
2131 if (record_type & PERF_RECORD_CPU)
2132 perf_output_put(&handle, cpu_entry);
2135 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
2137 if (record_type & PERF_RECORD_GROUP) {
2138 struct perf_counter *leader, *sub;
2139 u64 nr = counter->nr_siblings;
2141 perf_output_put(&handle, nr);
2143 leader = counter->group_leader;
2144 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2146 sub->pmu->read(sub);
2148 group_entry.event = sub->hw_event.config;
2149 group_entry.counter = atomic64_read(&sub->count);
2151 perf_output_put(&handle, group_entry);
2156 perf_output_copy(&handle, callchain, callchain_size);
2158 perf_output_end(&handle);
2165 struct perf_comm_event {
2166 struct task_struct *task;
2171 struct perf_event_header header;
2178 static void perf_counter_comm_output(struct perf_counter *counter,
2179 struct perf_comm_event *comm_event)
2181 struct perf_output_handle handle;
2182 int size = comm_event->event.header.size;
2183 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2188 perf_output_put(&handle, comm_event->event);
2189 perf_output_copy(&handle, comm_event->comm,
2190 comm_event->comm_size);
2191 perf_output_end(&handle);
2194 static int perf_counter_comm_match(struct perf_counter *counter,
2195 struct perf_comm_event *comm_event)
2197 if (counter->hw_event.comm &&
2198 comm_event->event.header.type == PERF_EVENT_COMM)
2204 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2205 struct perf_comm_event *comm_event)
2207 struct perf_counter *counter;
2209 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2213 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2214 if (perf_counter_comm_match(counter, comm_event))
2215 perf_counter_comm_output(counter, comm_event);
2220 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2222 struct perf_cpu_context *cpuctx;
2224 char *comm = comm_event->task->comm;
2226 size = ALIGN(strlen(comm)+1, sizeof(u64));
2228 comm_event->comm = comm;
2229 comm_event->comm_size = size;
2231 comm_event->event.header.size = sizeof(comm_event->event) + size;
2233 cpuctx = &get_cpu_var(perf_cpu_context);
2234 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2235 put_cpu_var(perf_cpu_context);
2237 perf_counter_comm_ctx(¤t->perf_counter_ctx, comm_event);
2240 void perf_counter_comm(struct task_struct *task)
2242 struct perf_comm_event comm_event;
2244 if (!atomic_read(&nr_comm_tracking))
2247 comm_event = (struct perf_comm_event){
2250 .header = { .type = PERF_EVENT_COMM, },
2251 .pid = task->group_leader->pid,
2256 perf_counter_comm_event(&comm_event);
2263 struct perf_mmap_event {
2269 struct perf_event_header header;
2279 static void perf_counter_mmap_output(struct perf_counter *counter,
2280 struct perf_mmap_event *mmap_event)
2282 struct perf_output_handle handle;
2283 int size = mmap_event->event.header.size;
2284 int ret = perf_output_begin(&handle, counter, size, 0, 0);
2289 perf_output_put(&handle, mmap_event->event);
2290 perf_output_copy(&handle, mmap_event->file_name,
2291 mmap_event->file_size);
2292 perf_output_end(&handle);
2295 static int perf_counter_mmap_match(struct perf_counter *counter,
2296 struct perf_mmap_event *mmap_event)
2298 if (counter->hw_event.mmap &&
2299 mmap_event->event.header.type == PERF_EVENT_MMAP)
2302 if (counter->hw_event.munmap &&
2303 mmap_event->event.header.type == PERF_EVENT_MUNMAP)
2309 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
2310 struct perf_mmap_event *mmap_event)
2312 struct perf_counter *counter;
2314 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2318 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2319 if (perf_counter_mmap_match(counter, mmap_event))
2320 perf_counter_mmap_output(counter, mmap_event);
2325 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
2327 struct perf_cpu_context *cpuctx;
2328 struct file *file = mmap_event->file;
2335 buf = kzalloc(PATH_MAX, GFP_KERNEL);
2337 name = strncpy(tmp, "//enomem", sizeof(tmp));
2340 name = d_path(&file->f_path, buf, PATH_MAX);
2342 name = strncpy(tmp, "//toolong", sizeof(tmp));
2346 name = strncpy(tmp, "//anon", sizeof(tmp));
2351 size = ALIGN(strlen(name)+1, sizeof(u64));
2353 mmap_event->file_name = name;
2354 mmap_event->file_size = size;
2356 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
2358 cpuctx = &get_cpu_var(perf_cpu_context);
2359 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
2360 put_cpu_var(perf_cpu_context);
2362 perf_counter_mmap_ctx(¤t->perf_counter_ctx, mmap_event);
2367 void perf_counter_mmap(unsigned long addr, unsigned long len,
2368 unsigned long pgoff, struct file *file)
2370 struct perf_mmap_event mmap_event;
2372 if (!atomic_read(&nr_mmap_tracking))
2375 mmap_event = (struct perf_mmap_event){
2378 .header = { .type = PERF_EVENT_MMAP, },
2379 .pid = current->group_leader->pid,
2380 .tid = current->pid,
2387 perf_counter_mmap_event(&mmap_event);
2390 void perf_counter_munmap(unsigned long addr, unsigned long len,
2391 unsigned long pgoff, struct file *file)
2393 struct perf_mmap_event mmap_event;
2395 if (!atomic_read(&nr_munmap_tracking))
2398 mmap_event = (struct perf_mmap_event){
2401 .header = { .type = PERF_EVENT_MUNMAP, },
2402 .pid = current->group_leader->pid,
2403 .tid = current->pid,
2410 perf_counter_mmap_event(&mmap_event);
2417 static void perf_log_period(struct perf_counter *counter, u64 period)
2419 struct perf_output_handle handle;
2423 struct perf_event_header header;
2428 .type = PERF_EVENT_PERIOD,
2430 .size = sizeof(freq_event),
2432 .time = sched_clock(),
2436 if (counter->hw.irq_period == period)
2439 ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
2443 perf_output_put(&handle, freq_event);
2444 perf_output_end(&handle);
2448 * Generic counter overflow handling.
2451 int perf_counter_overflow(struct perf_counter *counter,
2452 int nmi, struct pt_regs *regs, u64 addr)
2454 int events = atomic_read(&counter->event_limit);
2457 counter->hw.interrupts++;
2460 * XXX event_limit might not quite work as expected on inherited
2464 counter->pending_kill = POLL_IN;
2465 if (events && atomic_dec_and_test(&counter->event_limit)) {
2467 counter->pending_kill = POLL_HUP;
2469 counter->pending_disable = 1;
2470 perf_pending_queue(&counter->pending,
2471 perf_pending_counter);
2473 perf_counter_disable(counter);
2476 perf_counter_output(counter, nmi, regs, addr);
2481 * Generic software counter infrastructure
2484 static void perf_swcounter_update(struct perf_counter *counter)
2486 struct hw_perf_counter *hwc = &counter->hw;
2491 prev = atomic64_read(&hwc->prev_count);
2492 now = atomic64_read(&hwc->count);
2493 if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
2498 atomic64_add(delta, &counter->count);
2499 atomic64_sub(delta, &hwc->period_left);
2502 static void perf_swcounter_set_period(struct perf_counter *counter)
2504 struct hw_perf_counter *hwc = &counter->hw;
2505 s64 left = atomic64_read(&hwc->period_left);
2506 s64 period = hwc->irq_period;
2508 if (unlikely(left <= -period)) {
2510 atomic64_set(&hwc->period_left, left);
2513 if (unlikely(left <= 0)) {
2515 atomic64_add(period, &hwc->period_left);
2518 atomic64_set(&hwc->prev_count, -left);
2519 atomic64_set(&hwc->count, -left);
2522 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
2524 enum hrtimer_restart ret = HRTIMER_RESTART;
2525 struct perf_counter *counter;
2526 struct pt_regs *regs;
2529 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
2530 counter->pmu->read(counter);
2532 regs = get_irq_regs();
2534 * In case we exclude kernel IPs or are somehow not in interrupt
2535 * context, provide the next best thing, the user IP.
2537 if ((counter->hw_event.exclude_kernel || !regs) &&
2538 !counter->hw_event.exclude_user)
2539 regs = task_pt_regs(current);
2542 if (perf_counter_overflow(counter, 0, regs, 0))
2543 ret = HRTIMER_NORESTART;
2546 period = max_t(u64, 10000, counter->hw.irq_period);
2547 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
2552 static void perf_swcounter_overflow(struct perf_counter *counter,
2553 int nmi, struct pt_regs *regs, u64 addr)
2555 perf_swcounter_update(counter);
2556 perf_swcounter_set_period(counter);
2557 if (perf_counter_overflow(counter, nmi, regs, addr))
2558 /* soft-disable the counter */
2563 static int perf_swcounter_match(struct perf_counter *counter,
2564 enum perf_event_types type,
2565 u32 event, struct pt_regs *regs)
2567 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2570 if (perf_event_raw(&counter->hw_event))
2573 if (perf_event_type(&counter->hw_event) != type)
2576 if (perf_event_id(&counter->hw_event) != event)
2579 if (counter->hw_event.exclude_user && user_mode(regs))
2582 if (counter->hw_event.exclude_kernel && !user_mode(regs))
2588 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
2589 int nmi, struct pt_regs *regs, u64 addr)
2591 int neg = atomic64_add_negative(nr, &counter->hw.count);
2592 if (counter->hw.irq_period && !neg)
2593 perf_swcounter_overflow(counter, nmi, regs, addr);
2596 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
2597 enum perf_event_types type, u32 event,
2598 u64 nr, int nmi, struct pt_regs *regs,
2601 struct perf_counter *counter;
2603 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2607 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2608 if (perf_swcounter_match(counter, type, event, regs))
2609 perf_swcounter_add(counter, nr, nmi, regs, addr);
2614 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
2617 return &cpuctx->recursion[3];
2620 return &cpuctx->recursion[2];
2623 return &cpuctx->recursion[1];
2625 return &cpuctx->recursion[0];
2628 static void __perf_swcounter_event(enum perf_event_types type, u32 event,
2629 u64 nr, int nmi, struct pt_regs *regs,
2632 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
2633 int *recursion = perf_swcounter_recursion_context(cpuctx);
2641 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
2642 nr, nmi, regs, addr);
2643 if (cpuctx->task_ctx) {
2644 perf_swcounter_ctx_event(cpuctx->task_ctx, type, event,
2645 nr, nmi, regs, addr);
2652 put_cpu_var(perf_cpu_context);
2656 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
2658 __perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, regs, addr);
2661 static void perf_swcounter_read(struct perf_counter *counter)
2663 perf_swcounter_update(counter);
2666 static int perf_swcounter_enable(struct perf_counter *counter)
2668 perf_swcounter_set_period(counter);
2672 static void perf_swcounter_disable(struct perf_counter *counter)
2674 perf_swcounter_update(counter);
2677 static const struct pmu perf_ops_generic = {
2678 .enable = perf_swcounter_enable,
2679 .disable = perf_swcounter_disable,
2680 .read = perf_swcounter_read,
2684 * Software counter: cpu wall time clock
2687 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
2689 int cpu = raw_smp_processor_id();
2693 now = cpu_clock(cpu);
2694 prev = atomic64_read(&counter->hw.prev_count);
2695 atomic64_set(&counter->hw.prev_count, now);
2696 atomic64_add(now - prev, &counter->count);
2699 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
2701 struct hw_perf_counter *hwc = &counter->hw;
2702 int cpu = raw_smp_processor_id();
2704 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
2705 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2706 hwc->hrtimer.function = perf_swcounter_hrtimer;
2707 if (hwc->irq_period) {
2708 u64 period = max_t(u64, 10000, hwc->irq_period);
2709 __hrtimer_start_range_ns(&hwc->hrtimer,
2710 ns_to_ktime(period), 0,
2711 HRTIMER_MODE_REL, 0);
2717 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
2719 if (counter->hw.irq_period)
2720 hrtimer_cancel(&counter->hw.hrtimer);
2721 cpu_clock_perf_counter_update(counter);
2724 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
2726 cpu_clock_perf_counter_update(counter);
2729 static const struct pmu perf_ops_cpu_clock = {
2730 .enable = cpu_clock_perf_counter_enable,
2731 .disable = cpu_clock_perf_counter_disable,
2732 .read = cpu_clock_perf_counter_read,
2736 * Software counter: task time clock
2739 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
2744 prev = atomic64_xchg(&counter->hw.prev_count, now);
2746 atomic64_add(delta, &counter->count);
2749 static int task_clock_perf_counter_enable(struct perf_counter *counter)
2751 struct hw_perf_counter *hwc = &counter->hw;
2754 now = counter->ctx->time;
2756 atomic64_set(&hwc->prev_count, now);
2757 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2758 hwc->hrtimer.function = perf_swcounter_hrtimer;
2759 if (hwc->irq_period) {
2760 u64 period = max_t(u64, 10000, hwc->irq_period);
2761 __hrtimer_start_range_ns(&hwc->hrtimer,
2762 ns_to_ktime(period), 0,
2763 HRTIMER_MODE_REL, 0);
2769 static void task_clock_perf_counter_disable(struct perf_counter *counter)
2771 if (counter->hw.irq_period)
2772 hrtimer_cancel(&counter->hw.hrtimer);
2773 task_clock_perf_counter_update(counter, counter->ctx->time);
2777 static void task_clock_perf_counter_read(struct perf_counter *counter)
2782 update_context_time(counter->ctx);
2783 time = counter->ctx->time;
2785 u64 now = perf_clock();
2786 u64 delta = now - counter->ctx->timestamp;
2787 time = counter->ctx->time + delta;
2790 task_clock_perf_counter_update(counter, time);
2793 static const struct pmu perf_ops_task_clock = {
2794 .enable = task_clock_perf_counter_enable,
2795 .disable = task_clock_perf_counter_disable,
2796 .read = task_clock_perf_counter_read,
2800 * Software counter: cpu migrations
2803 static inline u64 get_cpu_migrations(struct perf_counter *counter)
2805 struct task_struct *curr = counter->ctx->task;
2808 return curr->se.nr_migrations;
2809 return cpu_nr_migrations(smp_processor_id());
2812 static void cpu_migrations_perf_counter_update(struct perf_counter *counter)
2817 prev = atomic64_read(&counter->hw.prev_count);
2818 now = get_cpu_migrations(counter);
2820 atomic64_set(&counter->hw.prev_count, now);
2824 atomic64_add(delta, &counter->count);
2827 static void cpu_migrations_perf_counter_read(struct perf_counter *counter)
2829 cpu_migrations_perf_counter_update(counter);
2832 static int cpu_migrations_perf_counter_enable(struct perf_counter *counter)
2834 if (counter->prev_state <= PERF_COUNTER_STATE_OFF)
2835 atomic64_set(&counter->hw.prev_count,
2836 get_cpu_migrations(counter));
2840 static void cpu_migrations_perf_counter_disable(struct perf_counter *counter)
2842 cpu_migrations_perf_counter_update(counter);
2845 static const struct pmu perf_ops_cpu_migrations = {
2846 .enable = cpu_migrations_perf_counter_enable,
2847 .disable = cpu_migrations_perf_counter_disable,
2848 .read = cpu_migrations_perf_counter_read,
2851 #ifdef CONFIG_EVENT_PROFILE
2852 void perf_tpcounter_event(int event_id)
2854 struct pt_regs *regs = get_irq_regs();
2857 regs = task_pt_regs(current);
2859 __perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, regs, 0);
2861 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
2863 extern int ftrace_profile_enable(int);
2864 extern void ftrace_profile_disable(int);
2866 static void tp_perf_counter_destroy(struct perf_counter *counter)
2868 ftrace_profile_disable(perf_event_id(&counter->hw_event));
2871 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2873 int event_id = perf_event_id(&counter->hw_event);
2876 ret = ftrace_profile_enable(event_id);
2880 counter->destroy = tp_perf_counter_destroy;
2881 counter->hw.irq_period = counter->hw_event.irq_period;
2883 return &perf_ops_generic;
2886 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
2892 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
2894 const struct pmu *pmu = NULL;
2897 * Software counters (currently) can't in general distinguish
2898 * between user, kernel and hypervisor events.
2899 * However, context switches and cpu migrations are considered
2900 * to be kernel events, and page faults are never hypervisor
2903 switch (perf_event_id(&counter->hw_event)) {
2904 case PERF_COUNT_CPU_CLOCK:
2905 pmu = &perf_ops_cpu_clock;
2908 case PERF_COUNT_TASK_CLOCK:
2910 * If the user instantiates this as a per-cpu counter,
2911 * use the cpu_clock counter instead.
2913 if (counter->ctx->task)
2914 pmu = &perf_ops_task_clock;
2916 pmu = &perf_ops_cpu_clock;
2919 case PERF_COUNT_PAGE_FAULTS:
2920 case PERF_COUNT_PAGE_FAULTS_MIN:
2921 case PERF_COUNT_PAGE_FAULTS_MAJ:
2922 case PERF_COUNT_CONTEXT_SWITCHES:
2923 pmu = &perf_ops_generic;
2925 case PERF_COUNT_CPU_MIGRATIONS:
2926 if (!counter->hw_event.exclude_kernel)
2927 pmu = &perf_ops_cpu_migrations;
2935 * Allocate and initialize a counter structure
2937 static struct perf_counter *
2938 perf_counter_alloc(struct perf_counter_hw_event *hw_event,
2940 struct perf_counter_context *ctx,
2941 struct perf_counter *group_leader,
2944 const struct pmu *pmu;
2945 struct perf_counter *counter;
2946 struct hw_perf_counter *hwc;
2949 counter = kzalloc(sizeof(*counter), gfpflags);
2951 return ERR_PTR(-ENOMEM);
2954 * Single counters are their own group leaders, with an
2955 * empty sibling list:
2958 group_leader = counter;
2960 mutex_init(&counter->mutex);
2961 INIT_LIST_HEAD(&counter->list_entry);
2962 INIT_LIST_HEAD(&counter->event_entry);
2963 INIT_LIST_HEAD(&counter->sibling_list);
2964 init_waitqueue_head(&counter->waitq);
2966 mutex_init(&counter->mmap_mutex);
2968 INIT_LIST_HEAD(&counter->child_list);
2971 counter->hw_event = *hw_event;
2972 counter->group_leader = group_leader;
2973 counter->pmu = NULL;
2976 counter->state = PERF_COUNTER_STATE_INACTIVE;
2977 if (hw_event->disabled)
2978 counter->state = PERF_COUNTER_STATE_OFF;
2983 if (hw_event->freq && hw_event->irq_freq)
2984 hwc->irq_period = div64_u64(TICK_NSEC, hw_event->irq_freq);
2986 hwc->irq_period = hw_event->irq_period;
2989 * we currently do not support PERF_RECORD_GROUP on inherited counters
2991 if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
2994 if (perf_event_raw(hw_event)) {
2995 pmu = hw_perf_counter_init(counter);
2999 switch (perf_event_type(hw_event)) {
3000 case PERF_TYPE_HARDWARE:
3001 pmu = hw_perf_counter_init(counter);
3004 case PERF_TYPE_SOFTWARE:
3005 pmu = sw_perf_counter_init(counter);
3008 case PERF_TYPE_TRACEPOINT:
3009 pmu = tp_perf_counter_init(counter);
3016 else if (IS_ERR(pmu))
3021 return ERR_PTR(err);
3026 atomic_inc(&nr_counters);
3027 if (counter->hw_event.mmap)
3028 atomic_inc(&nr_mmap_tracking);
3029 if (counter->hw_event.munmap)
3030 atomic_inc(&nr_munmap_tracking);
3031 if (counter->hw_event.comm)
3032 atomic_inc(&nr_comm_tracking);
3038 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3040 * @hw_event_uptr: event type attributes for monitoring/sampling
3043 * @group_fd: group leader counter fd
3045 SYSCALL_DEFINE5(perf_counter_open,
3046 const struct perf_counter_hw_event __user *, hw_event_uptr,
3047 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3049 struct perf_counter *counter, *group_leader;
3050 struct perf_counter_hw_event hw_event;
3051 struct perf_counter_context *ctx;
3052 struct file *counter_file = NULL;
3053 struct file *group_file = NULL;
3054 int fput_needed = 0;
3055 int fput_needed2 = 0;
3058 /* for future expandability... */
3062 if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
3066 * Get the target context (task or percpu):
3068 ctx = find_get_context(pid, cpu);
3070 return PTR_ERR(ctx);
3073 * Look up the group leader (we will attach this counter to it):
3075 group_leader = NULL;
3076 if (group_fd != -1) {
3078 group_file = fget_light(group_fd, &fput_needed);
3080 goto err_put_context;
3081 if (group_file->f_op != &perf_fops)
3082 goto err_put_context;
3084 group_leader = group_file->private_data;
3086 * Do not allow a recursive hierarchy (this new sibling
3087 * becoming part of another group-sibling):
3089 if (group_leader->group_leader != group_leader)
3090 goto err_put_context;
3092 * Do not allow to attach to a group in a different
3093 * task or CPU context:
3095 if (group_leader->ctx != ctx)
3096 goto err_put_context;
3098 * Only a group leader can be exclusive or pinned
3100 if (hw_event.exclusive || hw_event.pinned)
3101 goto err_put_context;
3104 counter = perf_counter_alloc(&hw_event, cpu, ctx, group_leader,
3106 ret = PTR_ERR(counter);
3107 if (IS_ERR(counter))
3108 goto err_put_context;
3110 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
3112 goto err_free_put_context;
3114 counter_file = fget_light(ret, &fput_needed2);
3116 goto err_free_put_context;
3118 counter->filp = counter_file;
3119 mutex_lock(&ctx->mutex);
3120 perf_install_in_context(ctx, counter, cpu);
3121 mutex_unlock(&ctx->mutex);
3123 fput_light(counter_file, fput_needed2);
3126 fput_light(group_file, fput_needed);
3130 err_free_put_context:
3140 * Initialize the perf_counter context in a task_struct:
3143 __perf_counter_init_context(struct perf_counter_context *ctx,
3144 struct task_struct *task)
3146 memset(ctx, 0, sizeof(*ctx));
3147 spin_lock_init(&ctx->lock);
3148 mutex_init(&ctx->mutex);
3149 INIT_LIST_HEAD(&ctx->counter_list);
3150 INIT_LIST_HEAD(&ctx->event_list);
3155 * inherit a counter from parent task to child task:
3157 static struct perf_counter *
3158 inherit_counter(struct perf_counter *parent_counter,
3159 struct task_struct *parent,
3160 struct perf_counter_context *parent_ctx,
3161 struct task_struct *child,
3162 struct perf_counter *group_leader,
3163 struct perf_counter_context *child_ctx)
3165 struct perf_counter *child_counter;
3168 * Instead of creating recursive hierarchies of counters,
3169 * we link inherited counters back to the original parent,
3170 * which has a filp for sure, which we use as the reference
3173 if (parent_counter->parent)
3174 parent_counter = parent_counter->parent;
3176 child_counter = perf_counter_alloc(&parent_counter->hw_event,
3177 parent_counter->cpu, child_ctx,
3178 group_leader, GFP_KERNEL);
3179 if (IS_ERR(child_counter))
3180 return child_counter;
3183 * Link it up in the child's context:
3185 child_counter->task = child;
3186 add_counter_to_ctx(child_counter, child_ctx);
3188 child_counter->parent = parent_counter;
3190 * inherit into child's child as well:
3192 child_counter->hw_event.inherit = 1;
3195 * Get a reference to the parent filp - we will fput it
3196 * when the child counter exits. This is safe to do because
3197 * we are in the parent and we know that the filp still
3198 * exists and has a nonzero count:
3200 atomic_long_inc(&parent_counter->filp->f_count);
3203 * Link this into the parent counter's child list
3205 mutex_lock(&parent_counter->mutex);
3206 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
3209 * Make the child state follow the state of the parent counter,
3210 * not its hw_event.disabled bit. We hold the parent's mutex,
3211 * so we won't race with perf_counter_{en,dis}able_family.
3213 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
3214 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
3216 child_counter->state = PERF_COUNTER_STATE_OFF;
3218 mutex_unlock(&parent_counter->mutex);
3220 return child_counter;
3223 static int inherit_group(struct perf_counter *parent_counter,
3224 struct task_struct *parent,
3225 struct perf_counter_context *parent_ctx,
3226 struct task_struct *child,
3227 struct perf_counter_context *child_ctx)
3229 struct perf_counter *leader;
3230 struct perf_counter *sub;
3231 struct perf_counter *child_ctr;
3233 leader = inherit_counter(parent_counter, parent, parent_ctx,
3234 child, NULL, child_ctx);
3236 return PTR_ERR(leader);
3237 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
3238 child_ctr = inherit_counter(sub, parent, parent_ctx,
3239 child, leader, child_ctx);
3240 if (IS_ERR(child_ctr))
3241 return PTR_ERR(child_ctr);
3246 static void sync_child_counter(struct perf_counter *child_counter,
3247 struct perf_counter *parent_counter)
3251 child_val = atomic64_read(&child_counter->count);
3254 * Add back the child's count to the parent's count:
3256 atomic64_add(child_val, &parent_counter->count);
3257 atomic64_add(child_counter->total_time_enabled,
3258 &parent_counter->child_total_time_enabled);
3259 atomic64_add(child_counter->total_time_running,
3260 &parent_counter->child_total_time_running);
3263 * Remove this counter from the parent's list
3265 mutex_lock(&parent_counter->mutex);
3266 list_del_init(&child_counter->child_list);
3267 mutex_unlock(&parent_counter->mutex);
3270 * Release the parent counter, if this was the last
3273 fput(parent_counter->filp);
3277 __perf_counter_exit_task(struct task_struct *child,
3278 struct perf_counter *child_counter,
3279 struct perf_counter_context *child_ctx)
3281 struct perf_counter *parent_counter;
3284 * If we do not self-reap then we have to wait for the
3285 * child task to unschedule (it will happen for sure),
3286 * so that its counter is at its final count. (This
3287 * condition triggers rarely - child tasks usually get
3288 * off their CPU before the parent has a chance to
3289 * get this far into the reaping action)
3291 if (child != current) {
3292 wait_task_inactive(child, 0);
3293 update_counter_times(child_counter);
3294 list_del_counter(child_counter, child_ctx);
3296 struct perf_cpu_context *cpuctx;
3297 unsigned long flags;
3300 * Disable and unlink this counter.
3302 * Be careful about zapping the list - IRQ/NMI context
3303 * could still be processing it:
3305 local_irq_save(flags);
3308 cpuctx = &__get_cpu_var(perf_cpu_context);
3310 group_sched_out(child_counter, cpuctx, child_ctx);
3311 update_counter_times(child_counter);
3313 list_del_counter(child_counter, child_ctx);
3316 local_irq_restore(flags);
3319 parent_counter = child_counter->parent;
3321 * It can happen that parent exits first, and has counters
3322 * that are still around due to the child reference. These
3323 * counters need to be zapped - but otherwise linger.
3325 if (parent_counter) {
3326 sync_child_counter(child_counter, parent_counter);
3327 free_counter(child_counter);
3332 * When a child task exits, feed back counter values to parent counters.
3334 * Note: we may be running in child context, but the PID is not hashed
3335 * anymore so new counters will not be added.
3337 void perf_counter_exit_task(struct task_struct *child)
3339 struct perf_counter *child_counter, *tmp;
3340 struct perf_counter_context *child_ctx;
3342 WARN_ON_ONCE(child != current);
3344 child_ctx = &child->perf_counter_ctx;
3346 if (likely(!child_ctx->nr_counters))
3350 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
3352 __perf_counter_exit_task(child, child_counter, child_ctx);
3355 * If the last counter was a group counter, it will have appended all
3356 * its siblings to the list, but we obtained 'tmp' before that which
3357 * will still point to the list head terminating the iteration.
3359 if (!list_empty(&child_ctx->counter_list))
3364 * Initialize the perf_counter context in task_struct
3366 void perf_counter_init_task(struct task_struct *child)
3368 struct perf_counter_context *child_ctx, *parent_ctx;
3369 struct perf_counter *counter;
3370 struct task_struct *parent = current;
3372 child_ctx = &child->perf_counter_ctx;
3373 parent_ctx = &parent->perf_counter_ctx;
3375 __perf_counter_init_context(child_ctx, child);
3378 * This is executed from the parent task context, so inherit
3379 * counters that have been marked for cloning:
3382 if (likely(!parent_ctx->nr_counters))
3386 * Lock the parent list. No need to lock the child - not PID
3387 * hashed yet and not running, so nobody can access it.
3389 mutex_lock(&parent_ctx->mutex);
3392 * We dont have to disable NMIs - we are only looking at
3393 * the list, not manipulating it:
3395 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
3396 if (counter != counter->group_leader)
3399 if (!counter->hw_event.inherit)
3402 if (inherit_group(counter, parent,
3403 parent_ctx, child, child_ctx))
3407 mutex_unlock(&parent_ctx->mutex);
3410 static void __cpuinit perf_counter_init_cpu(int cpu)
3412 struct perf_cpu_context *cpuctx;
3414 cpuctx = &per_cpu(perf_cpu_context, cpu);
3415 __perf_counter_init_context(&cpuctx->ctx, NULL);
3417 spin_lock(&perf_resource_lock);
3418 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
3419 spin_unlock(&perf_resource_lock);
3421 hw_perf_counter_setup(cpu);
3424 #ifdef CONFIG_HOTPLUG_CPU
3425 static void __perf_counter_exit_cpu(void *info)
3427 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3428 struct perf_counter_context *ctx = &cpuctx->ctx;
3429 struct perf_counter *counter, *tmp;
3431 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
3432 __perf_counter_remove_from_context(counter);
3434 static void perf_counter_exit_cpu(int cpu)
3436 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
3437 struct perf_counter_context *ctx = &cpuctx->ctx;
3439 mutex_lock(&ctx->mutex);
3440 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
3441 mutex_unlock(&ctx->mutex);
3444 static inline void perf_counter_exit_cpu(int cpu) { }
3447 static int __cpuinit
3448 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
3450 unsigned int cpu = (long)hcpu;
3454 case CPU_UP_PREPARE:
3455 case CPU_UP_PREPARE_FROZEN:
3456 perf_counter_init_cpu(cpu);
3459 case CPU_DOWN_PREPARE:
3460 case CPU_DOWN_PREPARE_FROZEN:
3461 perf_counter_exit_cpu(cpu);
3471 static struct notifier_block __cpuinitdata perf_cpu_nb = {
3472 .notifier_call = perf_cpu_notify,
3475 void __init perf_counter_init(void)
3477 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
3478 (void *)(long)smp_processor_id());
3479 register_cpu_notifier(&perf_cpu_nb);
3482 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
3484 return sprintf(buf, "%d\n", perf_reserved_percpu);
3488 perf_set_reserve_percpu(struct sysdev_class *class,
3492 struct perf_cpu_context *cpuctx;
3496 err = strict_strtoul(buf, 10, &val);
3499 if (val > perf_max_counters)
3502 spin_lock(&perf_resource_lock);
3503 perf_reserved_percpu = val;
3504 for_each_online_cpu(cpu) {
3505 cpuctx = &per_cpu(perf_cpu_context, cpu);
3506 spin_lock_irq(&cpuctx->ctx.lock);
3507 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
3508 perf_max_counters - perf_reserved_percpu);
3509 cpuctx->max_pertask = mpt;
3510 spin_unlock_irq(&cpuctx->ctx.lock);
3512 spin_unlock(&perf_resource_lock);
3517 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
3519 return sprintf(buf, "%d\n", perf_overcommit);
3523 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
3528 err = strict_strtoul(buf, 10, &val);
3534 spin_lock(&perf_resource_lock);
3535 perf_overcommit = val;
3536 spin_unlock(&perf_resource_lock);
3541 static SYSDEV_CLASS_ATTR(
3544 perf_show_reserve_percpu,
3545 perf_set_reserve_percpu
3548 static SYSDEV_CLASS_ATTR(
3551 perf_show_overcommit,
3555 static struct attribute *perfclass_attrs[] = {
3556 &attr_reserve_percpu.attr,
3557 &attr_overcommit.attr,
3561 static struct attribute_group perfclass_attr_group = {
3562 .attrs = perfclass_attrs,
3563 .name = "perf_counters",
3566 static int __init perf_counter_sysfs_init(void)
3568 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
3569 &perfclass_attr_group);
3571 device_initcall(perf_counter_sysfs_init);